Provenance-Aware Gap Imputation for Time Series Data

Preprocessing of time-series data is often driven by heuristic decisions that lack transparency and reproducibility, particularly when handling missing values and noise. In this work, we model preprocessing as a structured and traceable decision-making process. We propose a gap-aware strategy that adapts imputation methods based on the size of missing intervals. Depending on gap length, domain, and objective, missing values are imputed using domain-aware strategies that combine interpolation methods, polynomial fitting, and pattern-based reconstruction. Local signal bounds are applied where appropriate to avoid unrealistic signal behavior.

To improve transparency, we integrate preprocessing with provenance modeling by capturing decisions, parameters, and data-quality indicators as structured metadata. This enables reconstruction and analysis of preprocessing workflows. Our results demonstrate that the approach achieves competitive performance while significantly improving transparency and reproducibility.